Rather than only building another classifier, this thesis asks why detectors work when they do: which linguistic, statistical, or internal-representation signals actually drive the decision, and whether those signals are stable across model families and domains.

Tasks

• Probe internal representations (attention patterns, hidden-state features) of open-weight LLMs for signals correlated with detector decisions.
• Compare interpretable surface features (perplexity, burstiness, entropy, syntactic complexity) against learned classifier features via ablation/probing.
• Test whether the same features generalize across model families (e.g., a detector’s signal for one LLM’s output vs. another’s) and across domains (news, academic, social media).
• Produce a feature-attribution analysis explaining detector behavior, not just accuracy numbers.

Prerequisites

• Python, some background in interpretability/probing methods or willingness to learn
• strong analytical writing

Why this thesis is special

Moves beyond black-box benchmarking toward explaining detector behavior — useful both for building better detectors and for auditing their failure cases, with clear publication potential in interpretability-adjacent venues.